Thanks for the kind words!

I got lost the next half 😀

Thanks! :)

Thank you!

Thank you!

Great suggestion - will incorporate in newer lectures.

Good question - no not really that aren't multiple imputation. I mean lme will estimate the ones that make sense (i.e. if you have time 1 and 3, you'll get an idea of what two might have been like). There might even be something hidden in nlme that I'm not aware of.

Maybe! I think there could be a few reasons - repeated measures data is hard to tell sometimes without having more variables to test if that's the cause of the heterogeneity.

I've just realized that JASP added a MLM module. I haven't looked into it at all. I'm sure they did a great job implementing, but I'd probably vote for having more control over the analysis (I also feel this way about their SEM modules).

Here the nesting is within trial code - the response is the DV we are interested in. So, I would nest by time if you were interested in controlling for it that way.

@@StatisticsofDOOM Thanks! I guess I misunderstood this part. I think I'll watch it again!

Yep - everything should be here: github.com/doomlab/statsofdoom-files/tree/master/advanced/R/Other%20Files

Sure! You can control for that random effect as well. It would be another random intercept instead of a set of nested random intercepts like list(1|variable, 1|variable) rather than 1|var/var.

You can use any number greater than 2 really - just need enough variance in the chi-square distribution to get "randomness". 7 is just the number I picked a million years ago. And correlation wise...not really? I think you might be able to get the correlations out of the actual MLM part that could work pretty well (it shows when running lme with nlme package).

Thanks! You wouldn't happen to know to do any of these assumption tests in lmer instead of lme?

I believe you could do it approximately the same way - I haven't used lmer much, but the hard part would be getting the residuals out I would think. Looks like you can use resid(saved-model-fit) to get them. Additionally plot(saved-model-fit) will give you some diagnostic plots.

The concepts are the same but the code is not - you should use a glmer() model, so you can use a logistic regression. You can check out one of our papers that used this type of code osf.io/ex7js/ - we didn't do all the nested model testing, but you could. Notice the random effects are listed as (1|Variable) rather than in a special random argument.

@@StatisticsofDOOM I had the same question but I think this link doesn't work anymore.

@@ahmetaltnok3572 Unfortunately, the autoformatting on youtube included the ) - the link is osf.io/ex7js/

Yes, I've heard you can do that as well. I don't know if they are equivalent in the sense that they will always give you the same answer, but they are likely good options to demonstrate the need to control for correlated error. I often will still use random intercept even if it's not significant just because it seems odd to NOT control for participant correlated error when using repeated measures designs.

Thank you Dr. B. Much love from another MSU (Michigan State).

hi there i am a phd student do you master multilevel model

Yes! That's a good reason as well :) then you are using regression and controlling for correlated error.

You are correct, and I believe it's mostly to help with interpretation of the slopes as well (i.e., mean centering allows you to know that the main regression is the average level for slopes).

Statistics of DOOM Thank you for your reply, Professor. I much appreciate it. I am not quite sure that I follow your answer though. The average of random effects are zero both when centering and not centering variables. In both case the random effects are spread around the fixed effect, and do so with mean zero. So whether or not centering is done, shouldn't the interpretation be the same for the slopes? (Given that they are centered around the grand mean, that is, and not around subject level means.)

Oh right, I was talking about the fixed effect interpretation not the random effects - sorry about that! You are correct.

The short answer is there isn't a good answer? There's a couple new papers that talk about this idea: largescaleassessmentsineducation.springeropen.com/articles/10.1186/s40536-018-0061-2

www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=2&cad=rja&uact=8&ved=2ahUKEwjj9Zrko6feAhVmtIMKHRm9BlgQFjABegQICBAC&url=http%3A%2F%2Fwww.psy.mq.edu.au%2Fpsystat%2Fdocuments%2Fstandardised_effect_size_in_mixed_ML_models.pdf&usg=AOvVaw3y89_kkv8uLxjPovhv11uv is another one.

I’ll add it to the list! Bit swamped right now unfortunately 😖

I tried an interception model using school number as a level 2 variable, the ICC of this model is 0.16, indicating somehow similarity within a school, but I am not sure whether my small sample sizes at both level 2 and level 1 are suitable for multilevel modeling

Yes! You just need N > 1 observations in the nesting variables.

The intercept only model does not control for the repeated measures component, so you can use lm() or gls() - I just happen to like gls() in this case.

Especially because of the correlated error terms - check this out: socialsciences.mcmaster.ca/jfox/Books/Companion/appendices/Appendix-Timeseries-Regression.pdf

Quick reply right there :D got it thanks

Always good to know what’s missing where - I often catch coding errors or other issues by screening.

Actually have done a bunch of these recently but have not made a video of them. You can extend the use mlm to log using the lme4 package the glmer() function: glmer(DV ~ IV + (1|part_no), data = data, family = binomial(), control = glmerControl(optimizer = "bobyqa"), nAGQ = 1) The family/linking function converts it to log regression.

No to machine learning, as that's not my area.

Hi, I really appreciate your guidance.Thanks for that. I am dealing with an insurance company. I have to predict when a policy will lapse (or churn) based on the details that are sitting at lower levels (at individual clients in a policy, and then at different covers (riders) that each client might have). So there are three levels in total, which are a Policy, a Client, and a Cover level. In most MLM examples that I have seen, the highest level groups usually have many observations e.g. students in classrooms or workers in work groups. Here, if I say, each Policy is a group then there are usually less than 10 observations in each group. Is this OK to think a policy as the highest level group? Especially we want to predict whether a policy will lapse or not. Please see the picture of the data sample here: pin.it/5ksmOdj Kind Regards, Umair

Yes, I believe it would be ok to group by that factor - if it doesn't run then you could try a different combination.

How odd - I definitely recorded it full screen. :| the code is provided on our companion website to view though, if you need to see it completely.

@@StatisticsofDOOM No website anymore?

@@sallyletsinger3789 It's there, just need to update all the links and finish putting together the new layout: statisticsofdoom.com/page/advanced-statistics/

Are you wanting time to be the random factor? Or a fixed effect? Not completely following.

@@StatisticsofDOOM I am tracking hormones throughout the menstrual cycle at three time points and interested if they affect behavior. If I want to know if within change in hormones affects behavior, would i have to include time into the model? Hope that is clear.

@@rayray0313 Ok, so if you expect some sort of linear change over time, I would put time in equation - if you expect it to be variable, I would suggest some of the level centering techniques that I see clinical people do - here's what I think is an open access article on the topic: www.ncbi.nlm.nih.gov/pmc/articles/PMC3059070/

@@StatisticsofDOOM This was very helpful. Thank you so much! Will you do any videos on multi-level modeling in JAMOVI?

@@rayray0313 ohhh! I didn't know jamovi did that - I will check it out and add it to my list.

1) if you have an interaction you are interested in breaking down, you could do simple effects 2) you could run pairwise t-tests if you are only controlling for random intercepts by participant - I might suggest a simple effects analysis (split one variable into separate datasets, run the same MLM on the each dataset separately without that split variable in the equation). That would hold all things constant, in case your model gets more complex.